Executive brief
CVAT is an open-source tool for annotating video and image datasets used in computer vision projects. The audio annotation feature allows users to create guides in Markdown format; however, malicious JavaScript code embedded in these guides is not sanitized before display. An attacker with guide-creation privileges can inject code that executes in the browsers of other users viewing the guide, potentially allowing them to perform unauthorized actions on behalf of those users.
Technical details
This is a stored cross-site scripting (XSS) vulnerability in the audio-task annotation guide renderer located in cvat-ui/src/audio/components/annotation-page/audio-workspace/top-bar/audio-right-group.tsx. The vulnerable component passes user-supplied Markdown guide content to MDEditor without applying the rehype-sanitize plugin, which is responsible for removing malicious HTML and JavaScript. An authenticated user with permissions to create or edit annotation guides can store malicious JavaScript code that executes when any other user opens the guide, allowing arbitrary CVAT API requests to be made with the victim's session privileges. The vulnerability affects versions 2.68.0 through 2.69.x and is resolved in version 2.70.0.
Affected products
- CVAT AI CVAT 2.68.0 through 2.69.x
Timeline
- 2026-08-20: disclosed
- 2026-07-14: patched: Fixed in version 2.70.0